# Regional Prompter: different prompts for different parts of one image > An AUTOMATIC1111 extension that divides the canvas into regions and prompts each separately, either in latent space or by masking attention inside the U-Net. It supports Forge and reForge, has no releases of its own, and its support matrix is more honest than most. **hako-mikan/sd-webui-regional-prompter** — set prompt to divided region - Repository: https://github.com/hako-mikan/sd-webui-regional-prompter - Stars: 1,815 · Forks: 152 - Language: Python - License: AGPL-3.0 - Published: 2026-10-16 · Updated: 2026-10-16 · Language: en - Canonical page: https://hysenlabs.com/projects/hako-mikan-sd-webui-regional-prompter ## The idea is per prompt work inside the U-Net, not per prompt copies of it The repository description is one line: set prompt to divided region. The overview section is where the mechanism becomes clear. The README contrasts this with Latent Couple, which performs U-Net calculations on a per prompt basis, meaning the whole network runs once per prompt. Regional Prompter instead performs per prompt calculations inside the U-Net, so the model runs once and the regions are separated during that single pass. Latent mode is also supported as an alternative path. The credit line is specific as well, thanking furusu for initiating the idea and linking to a Japanese write-up of the details. That distinction is the whole engineering argument. Running the U-Net once per region is simple and produces results that read as a collage, because each half composes for itself. Doing the separation inside one pass is what keeps an image token seeing the whole image while only seeing its own region's prompt, which is the property that makes the output one picture rather than a mosaic. The extension is a custom script for AUTOMATIC1111's stable-diffusion-webui, and the README is bilingual with a separate README.JP.md, plus differential and prompt documents in English and Japanese. ## BREAK is the separator, and the settings decide how regions are sized The usage example is short enough to memorise. The prompt is a single line per region separated by BREAK: ``` green hair twintail BREAK red blouse BREAK blue skirt ``` The settings block that produces the described image is: ``` Active : On Use base prompt : Off Divide mode : Vertical Divide Ratio : 1,1,1 Base Ratio : ``` That divides the image vertically into three equal parts and applies the three prompts from top to bottom. Several details matter more than they look. Active must be toggled or the extension does nothing. Negative prompts are separated by BREAK as well, and if you do not enter BREAK for the negatives, the same negative applies to every region. Using ADDROW or ADDCOL anywhere in the prompt automatically switches on 2D region mode, and a semicolon separator in the divide ratio does the same thing. Divide ratio accepts weights, fractions or pixel values. Entering 1,1,1 divides into thirds, 3,1,1 gives 60, 20 and 20 percent, and 0.1,0.1,0.1 is equivalent to 1,1,1. For accuracy the README recommends entering pixel values corresponding to the height or width, so 300,100,112 maps onto a 512 image. Base ratio controls how much of the base prompt survives in each region. A value of 0.2 produces 20 percent base and 80 percent region, and it can be set per region as 0.2,0.3,0.5 or applied uniformly from a single value. With the base prompt option enabled, the first BREAK separated prompt becomes the base prompt, which means one extra segment is required. ## The compatibility tables are the most useful thing in the README Two tables, dated September 2026 and January 2025, tell you exactly what works and where. The newer one covers Forge Neo models. Z-Image, Krea2 and Anima are supported in Latent mode, Attention mode is marked as only partly working for Z-Image and supported for the other two, and Region LoRA is not supported for any of them. The symbols carry real meaning. Full support is a circle. The triangle means scenery and figures separate well, but two regions that each want to be the single centred subject on a plain background tend to merge into one. That is the honest failure mode stated in the open. The reason given for Region LoRA breaking is architectural: it still names blocks the way the U-Nets do, so on these models it has nothing to match on and does nothing. The second reason is that these models do not cut the prompt into 75 token chunks, so Attention mode cannot divide one prompt the way it does on SD and SDXL. Each region is encoded on its own instead, and the separation happens inside the model's own attention. Anima is described separately. Rather than having a single softmax to separate regions inside, it has its own cross attention. That cross attention is blocked in the same way, while the self attention is leaned on with a small penalty rather than fully closed, and only for the first third of the schedule, after which nothing is masked and the model knits the regions together. The README is explicit about the alternative: closing the self attention separates harder but turns the picture into a collage. ## Attention mode has a cost, and the README names the reason For two regions, Attention mode costs about the same as Latent mode on those Forge Neo models, and it gets cheaper relative to Latent as regions are added, because the model runs once rather than once per region. That is a fair argument for the default. It is still slower than a plain generation, and the reason is specific enough to be useful. A masked attention has to go through PyTorch's own kernel, because SageAttention and the other fused kernels accept a mask argument and then ignore it. That is the kind of detail that usually goes undocumented. If you are running SageAttention for speed, switching this extension to Attention mode silently loses the benefit, and the extension is telling you why. The older table covers clients rather than models, and it has the same candour. Attention mode works on A1111, Forge and reForge. Latent mode is fully supported on A1111 and only partly on Forge and reForge, with the note that it may not function properly with large batch sizes and that it depends on VRAM. LoRA in Latent mode works on A1111 and Forge but not reForge. The README also explains the LoRA versus LoCon and LoHa split, which comes from Web-UI limitations rather than from this extension. Plain LoRA can be applied with no speed decrease. LoCon and LoHa work when the Use LoHa or other option is enabled, at the cost of slower generation, because the Web-UI cannot perform the specific optimisations needed and does not support mid strength changes. Forge has no such restriction. ## No releases, no version tags, and an AGPL extension The repository has no GitHub releases. The last push was on 2026-09-04, which matches the newest update entry in the README dated 2026.09.04 in Japan Standard Time, and the repository is not archived. Installation is therefore the standard AUTOMATIC1111 extension route rather than a documented command, and the README does not reproduce one in the text available here. The practical instruction is to install it as a custom script for the web UI from the repository. The file layout reflects how an A1111 extension is built. There is a scripts/ directory holding the extension entry point, regional_masks/ for the mask generation, javascript/ for the interface, style.css for the panel styling, and .github/ for the workflow. LICENCE sits at the root. The licence is AGPL-3.0, which is the strongest copyleft of the common ones and is worth thinking about before you bundle this into anything you distribute. This is not legal advice; the point is only that AGPL obligations reach further than MIT or Apache when code is offered over a network. The README is maintained in both English and Japanese with parallel documentation files, and the translation is complete enough that the feature list, the mode tables and the usage section all appear in both. For an extension maintained by one author, that is more maintenance effort than the project size would predict. The repository topics list custom script, region, prompter and stable diffusion webui, which is a fairly literal description of a tool with exactly one job. ## Where it is the wrong tool If you need two independent subjects centred in their own halves of the frame, the README's own warning applies: that is exactly the case where regions merge or the result becomes a collage. Two regions each wanting to be the subject on a plain background is the failure it does not solve. If your base model is one of the newer Forge Neo models and you rely on Region LoRA, it does nothing there. That is not a quality degradation, it is a feature that silently has no effect. If you are running SageAttention for throughput, the masked attention path will not use it. That trade may be worth it, but it should be a decision. If you need reproducible prompts across a batch, note that Latent mode on Forge and reForge is flagged as depending on VRAM and as possibly failing with large batch sizes, so the same settings may not produce the same thing at a different batch size. And if you are not on A1111, Forge or reForge, this extension does not apply at all. The README's compatibility tables cover three clients and nothing else, which is a narrow but clearly drawn line. ## Conclusion Adopt Regional Prompter if you run AUTOMATIC1111, Forge or reForge and you need a character in one part of the frame and a setting in another, since per region prompting is something the base UI does not offer and the BREAK syntax is trivial to learn. Do not adopt it expecting uniform behaviour across models, because the README's own support matrix marks Region LoRA as unsupported on Z-Image, Krea2 and Anima and Latent mode as only partly working on Z-Image. Before enabling anything beyond the default, read the mode and client compatibility tables at the top of the README, and expect the masked attention path to cost more time than a plain generation because fused kernels like SageAttention ignore the mask. ## FAQ ### How do I use the Regional Prompter extension? Toggle Active, then separate prompts with the BREAK keyword. Divide ratio sets the regions, so 1,1,1 gives equal thirds and 3,1,1 gives 60, 20 and 20 percent, and ADDROW or ADDCOL in the prompt switches on 2D region mode. ### What is the difference between Latent mode and Attention mode? Regional Prompter performs per-prompt calculations inside U-Net, unlike Latent Couple, which performs U-Net calculations per prompt. Attention mode can keep the result one picture rather than a collage, but it is still slower than plain generation because masked attention cannot use fused kernels such as SageAttention. ### Which web UI clients are supported? A1111, Forge and reForge for Attention mode. Latent mode is fully supported on A1111 but only partly on Forge and reForge, where it may not work properly with large batch sizes and depends on VRAM. ### Does Regional Prompter work with Z-Image, Krea2 and Anima? Yes, with limits. Latent mode is supported and Attention mode is supported for Krea2 and Anima but only partly for Z-Image, while Region LoRA is not supported on any of them because it names blocks the way the U-Nets do. ### Can I apply one base prompt plus region prompts? Yes. Enable the base prompt option and the first BREAK separated prompt becomes the base prompt, so one extra segment is needed. The Base ratio controls the mix, with 0.2 meaning 20 percent base prompt and 80 percent region prompt, and it can be set per region. ## Sources - [hako-mikan/sd-webui-regional-prompter on GitHub](https://github.com/hako-mikan/sd-webui-regional-prompter) - [Issues](https://github.com/hako-mikan/sd-webui-regional-prompter/issues) - [License: AGPL-3.0](https://github.com/hako-mikan/sd-webui-regional-prompter/blob/main/LICENSE) - [README](https://github.com/hako-mikan/sd-webui-regional-prompter/blob/main/README.md) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/hako-mikan-sd-webui-regional-prompter